Claude ID Verification, GLM 5.2 Self-Host, Codex UI Flip — Jun 21
435 messages · 62 active members
Overview
Topics
@arielletolome surfaced Anthropic's new identity verification policy, prompting debate about account farming workarounds and worries it won't be limited to US nationals. @jcartu and @arielletolome speculated the move is downstream of government pressure after Fable's capabilities scared regulators, predicting a domino effect across other providers.
@jcartu shared a HuggingFace release of GLM-5.2-504B-FullKD and made a case for buying 4-8 RTX 6000 Pros to self-host: ~0 KLD loss, 8 concurrent sessions at 60-80tps, near Opus 4.8 coding quality. Card prices jumped 750k→850k rub in 3 months. Tradeoffs: NVFP4 quant degradation, 120k ctx at 70tps on TP4 vs 262k ctx at 40tps on DCP4.
Builders reported a flip — Codex/GPT-5.5 now wins on UI generation and copywriting, while Opus is relegated to strategy and creative writing. @tounano shared the standout workflow: generate UI mockups (including hero shots) with GPT-Image-2, then feed screenshots to Codex for ~90% accurate code. Iterating on images beats iterating on code and parallelizes across chat windows.
@bartekadamczyk asked how builders orchestrate multiple agents and manage workflow memory. @sibunting runs one main agent in Telegram delegating to others via group chat, with Jira ticketing, GH Actions CI gates, and a custom control center for memory/context/billing. @jaypozo uses a Telegram ops agent with a GitHub Issues task repo. @navuud prefers Claude tmux dispatched to Telegram.
@iamgalba detailed his Google Ads scaling: 20-30 presell pages/month across listicles, advertorials, how-tos, and VSLs, each with 20 PMax image ads and A/B/C PostHog tests; winners promoted to Demand Gen on Discover + YT in-feed. Separately, @mb29266 and @Uziarbiv discussed deepfake celebrity advertorials on Taboola/CBS placements causing CTR inflation legitimate buyers can't compete with.
Key Takeaways
- Claude ID verification is rolling out — widely seen as government-driven and likely to cascade to other providers as they release more powerful models.
- Self-hosted GLM 5.2 on 4-8 RTX 6000 Pros delivers near-Opus 4.8 coding at electricity cost; card prices climbing fast (750k→850k rub in 3 months).
- New UI stack: generate mockups with GPT-Image-2 first, then have Codex convert screenshots to code — faster and more parallelizable than prompting code directly.
- Working agent orchestration: one ops agent in Telegram delegating to specialists, GitHub Issues or Jira for task tracking, custom control center for memory/context/billing.
- For AI trading or any prompted system, the evaluation harness for accepting/rejecting outputs matters more than the model or the prompt itself.
Hot Threads
Claude ID verification announcement and implications
GPT-Image-2 + Codex as the new UI generation stack
Deepfake celeb advertorials inflating CTRs across native